phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-20 04:56:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

implement embedding generation in supported inference providers (#589)

Browse source 
This PR adds the ability to generate embeddings in all supported
inference providers.

```
pytest -v -s llama_stack/providers/tests/inference/test_embeddings.py -k "bedrock" --inference-model="amazon.titan-embed-text-v2:0"  --env EMBEDDING_DIMENSION=1024

 pytest -v -s -k "vllm"  --inferrence-model="intfloat/e5-mistral-7b-instruct"  llama_stack/providers/tests/inference/test_embeddings.py --env EMBEDDING_DIMENSION=4096  --env VLLM_URL="http://localhost:9798/v1"

pytest -v -s --inference-model="nomic-ai/nomic-embed-text-v1.5"  llama_stack/providers/tests/inference/test_embeddings.py  -k "fireworks"  --env FIREWORKS_API_KEY=<API_KEY>--env EMBEDDING_DIMENSION=128

pytest -v -s --inference-model="togethercomputer/m2-bert-80M-2k-retrieval"  llama_stack/providers/tests/inference/test_embeddings.py  -k "together"  --env TOGETHER_API_KEY=<API_KEY>--env EMBEDDING_DIMENSION=768

pytest -v -s -k "ollama"  --inference-model="all-minilm:v8"  llama_stack/providers/tests/inference/test_embeddings.py --env EMBEDDING_DIMENSION=384

 torchrun $CONDA_PREFIX/bin/pytest -v -s -k "meta_reference" --inference-model="sentence-transformers/all-MiniLM-L6-v2"  llama_stack/providers/tests/inference/test_embeddings.py --env EMBEDDING_DIMENSION=384

```
This commit is contained in:
Dinesh Yeduguru 2024-12-12 11:17:39 -08:00
parent 6a23f24ee0
commit d362d2d740
32 changed files with 597 additions and 143 deletions
Download patch file Download diff file Expand all files Collapse all files

8
llama_stack/providers/remote/memory/pgvector/__init__.py
View file

@ -4,12 +4,16 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Dict
from llama_stack.providers.datatypes import Api, ProviderSpec
from .config import PGVectorConfig
async def get_adapter_impl(config: PGVectorConfig, _deps):
async def get_adapter_impl(config: PGVectorConfig, deps: Dict[Api, ProviderSpec]):
from .pgvector import PGVectorMemoryAdapter
impl = PGVectorMemoryAdapter(config)
impl = PGVectorMemoryAdapter(config, deps[Api.inference])
await impl.initialize()
return impl

38
llama_stack/providers/remote/memory/pgvector/pgvector.py
View file

@ -16,9 +16,9 @@ from pydantic import BaseModel, parse_obj_as
from llama_stack.apis.memory import * # noqa: F403
from llama_stack.providers.datatypes import MemoryBanksProtocolPrivate
from llama_stack.providers.datatypes import Api, MemoryBanksProtocolPrivate
from llama_stack.providers.utils.memory.vector_store import (
ALL_MINILM_L6_V2_DIMENSION,
BankWithIndex,
EmbeddingIndex,
)
@ -120,8 +120,9 @@ class PGVectorIndex(EmbeddingIndex):
class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
def __init__(self, config: PGVectorConfig) -> None:
def __init__(self, config: PGVectorConfig, inference_api: Api.inference) -> None:
self.config = config
self.inference_api = inference_api
self.cursor = None
self.conn = None
self.cache = {}
@ -160,27 +161,17 @@ class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
async def shutdown(self) -> None:
pass
async def register_memory_bank(
self,
memory_bank: MemoryBank,
) -> None:
async def register_memory_bank(self, memory_bank: MemoryBank) -> None:
assert (
memory_bank.memory_bank_type == MemoryBankType.vector.value
), f"Only vector banks are supported {memory_bank.memory_bank_type}"
upsert_models(
self.cursor,
[
(memory_bank.identifier, memory_bank),
],
upsert_models(self.cursor, [(memory_bank.identifier, memory_bank)])
index = PGVectorIndex(memory_bank, memory_bank.embedding_dimension, self.cursor)
self.cache[memory_bank.identifier] = BankWithIndex(
memory_bank, index, self.inference_api
)
index = BankWithIndex(
bank=memory_bank,
index=PGVectorIndex(memory_bank, ALL_MINILM_L6_V2_DIMENSION, self.cursor),
)
self.cache[memory_bank.identifier] = index
async def unregister_memory_bank(self, memory_bank_id: str) -> None:
await self.cache[memory_bank_id].index.delete()
del self.cache[memory_bank_id]
@ -203,14 +194,13 @@ class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
index = await self._get_and_cache_bank_index(bank_id)
return await index.query_documents(query, params)
self.inference_api = inference_api
async def _get_and_cache_bank_index(self, bank_id: str) -> BankWithIndex:
if bank_id in self.cache:
return self.cache[bank_id]
bank = await self.memory_bank_store.get_memory_bank(bank_id)
index = BankWithIndex(
bank=bank,
index=PGVectorIndex(bank, ALL_MINILM_L6_V2_DIMENSION, self.cursor),
)
self.cache[bank_id] = index
return index
index = PGVectorIndex(bank, bank.embedding_dimension, self.cursor)
self.cache[bank_id] = BankWithIndex(bank, index, self.inference_api)
return self.cache[bank_id]